Class: ToyLMCuda

Inherits:
Object
  • Object
show all
Defined in:
lib/toy/models/transformer_lm_cuda.rb

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(arch) ⇒ ToyLMCuda

Returns a new instance of ToyLMCuda.



35
36
37
38
39
40
41
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 35

def initialize(arch)
  @arch    = arch
  @max_T   = 256
  @kv      = nil
  @gguf_handle = nil
  @loaded  = false
end

Instance Attribute Details

#arch ⇒ Object (readonly)

Returns the value of attribute arch.



33
34
35
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 33

def arch
  @arch
end

#max_T ⇒ Object

Returns the value of attribute max_T.



33
34
35
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 33

def max_T
  @max_T
end

#tokenizer ⇒ Object (readonly)

Returns the value of attribute tokenizer.



33
34
35
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 33

def tokenizer
  @tokenizer
end

Instance Method Details

#decode_step(token_id, pos) ⇒ Object 



108
109
110
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 108

def decode_step(token_id, pos)
  SmolLM2KVCuda.decode_step(@kv, token_id, pos)
end

#generate(prompt_ids, n_new, sampler_config = nil) ⇒ Object 



112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 112

def generate(prompt_ids, n_new, sampler_config = nil)
  if !@loaded
    puts "ToyLMCuda.generate: model not loaded; call .load(path) first"
    return prompt_ids
  end
  ids = []
  j = 0
  while j < prompt_ids.length
    ids.push(prompt_ids[j])
    j = j + 1
  end

  i = 0
  while i < prompt_ids.length
    decode_step(prompt_ids[i], i)
    i = i + 1
  end

  ctx = nil
  if sampler_config != nil
    ctx = SamplerContext.new(ids, sampler_config.seed)
  end

  n = 0
  while n < n_new
    pos = ids.length
    last_id = ids[pos - 1]
    logits = decode_step(last_id, pos)
    pick = -1
    if sampler_config == nil
      pick = Sampler.argmax(logits)
    else
      logits = Sampler.repetition_penalty(logits, ctx, sampler_config.rep_penalty)
      logits = Sampler.temperature(logits, sampler_config.temperature)
      logits = Sampler.top_k(logits, sampler_config.top_k)
      logits = Sampler.top_p(logits, sampler_config.top_p)
      pick   = Sampler.pick(logits, sampler_config, ctx)
      ctx.generated_ids.push(pick)
    end
    ids.push(pick)
    if pick == @arch.eos_id
      break
    end
    n = n + 1
  end
  ids
end

#load(path) ⇒ Object 



47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
 
# File 'lib/toy/models/transformer_lm_cuda.rb', line 47

def load(path)
  flags = GGUFLoad.detect_smollm2_flags(path)
  cfg   = SmolLM2ConfigLoader.read(path)

  probe = TinyNNCuda.tnn_gguf_load(path)
  is_native = false
  if probe != nil
    is_native = (TinyNNCuda.tnn_gguf_get_bool(probe, "toy.ggml_native") == 1)
    TinyNNCuda.tnn_gguf_free(probe)
  end

  kv = SmolLM2KVFFICacheCuda.new
  # P5.1: KV_Q8=1 opts into Q8_0 storage for the K cache. See the CPU
  # mirror in lib/transformer_lm.rb for the full rationale.
  if (ENV["KV_Q8"] || "") == "1"
    kv.enable_kv_q8!
  end
  # P4.1: FLASH_ATTN=1 → ggml_flash_attn_ext for attention.
  if (ENV["FLASH_ATTN"] || "") == "1"
    kv.enable_flash_attn!
  end
  # #76 fix: wire QK-norm through to the engine, parity with
  # load_cpu in lib/toy/models/transformer_lm.rb. This call site
  # previously passed only 5 of realize_for_mmap's 6 args; Spinel
  # zero-fills missing call args WITHOUT a diagnostic, so qk_norm
  # arrived as false and Qwen3's per-head Q/K RMS-norms were never
  # built on CUDA → degenerate decode (CPU was coherent).
  # 1 = Qwen3-style ([d_head] shared), 2 = OLMoE/Granite-style
  # ([d_model] packed, per-head sliced gamma). Must be set BEFORE
  # realize_for_mmap.
  kv.qk_norm_kind = flags.qk_norm_kind
  qk_norm_on = flags.qk_norm
  # NO_QK_NORM=1 turns the norm off entirely as a diagnostic
  # (same env knob as the CPU loader).
  if (ENV["NO_QK_NORM"] || "") == "1"
    qk_norm_on = false
    kv.qk_norm_kind = 0
  end

  if is_native
    @gguf_handle = TinyNNCuda.tnn_gguf_load(path)
    kv.realize_for_mmap(@gguf_handle, cfg, @max_T, flags.untied, flags.qkv_bias, qk_norm_on)
  else
    if qk_norm_on
      # Fail loud (never mask): the legacy copy-load path has no
      # QK-norm support — realize_for can't allocate the gamma
      # tensors, so decode would be silently degenerate.
      puts "ToyLMCuda.load: " + path + " needs QK-norm but is not in " +
           "toy.ggml_native layout; the legacy copy-load path cannot " +
           "apply QK-norm. Re-convert with --ggml-native. Aborting."
      exit 1
    end
    kv.realize_for(@max_T, cfg.d_model, cfg.d_ff, cfg.n_heads, cfg.n_kv,
                   cfg.n_layers, cfg.vocab, cfg.rope_base, cfg.rms_eps,
                   flags.untied, flags.qkv_bias)
    GGUFLoad.load_kv_cache_auto(kv, path)
  end
  @kv = kv
  @loaded = true
end